Overview

An HR director asks AI to draft an email announcing a new work-from-home policy. They read it, it sounds fine, they send it. Later, an employee with mobility issues who relies on WFH says the message made no mention of how the new policy would be administered, whether exceptions could be requested, how accommodations work, whether she's affected. The email created more questions and frustration.

Three miles away, another HR director asks AI to draft a meeting agenda. The output is clear, structured, prioritized. They send it, meeting runs smoothly, they're done in 45 minutes instead of the usual 60. Same tool, wildly different outcomes.

The difference isn't the tool. It's knowing which tasks are safe to delegate to AI and which ones demand human judgment. This lesson teaches you to build your personal trust matrix: which HR tasks can you hand mostly to AI, which need heavy editing, which do you never use AI for.

Why This Matters for HR Professionals

You're drowning in work. If you could safely delegate 30% of your writing and drafting to AI, you'd get back 5-10 hours a week. That's real. But it only works if you know the risk level of each task.

Some HR tasks have low risk. An AI agenda, a recruiting email draft, a FAQ, if AI gets it 80% right, you're ahead. You'll edit it anyway. You're saving the starting-from-scratch time.

Other tasks are high-risk. A PIP document. A termination letter. An accommodation response. An investigation summary. These might end up in court. "AI drafted it" isn't a defense. You're liable for what goes out with your signature.

Your job is to know the difference. To build a matrix that says: "For this category of task, AI is a time-saver. For this one, AI is a starting point that needs heavy oversight. For this one, I'm not using AI."

The Trust Decision Framework

There are three variables:

1. Legal risk: Could this create liability if it's wrong? Is there compliance, ADA, anti-discrimination, or contract law exposure?

2. Accuracy requirements: How much does correctness matter? Is this about facts that can be verified, or judgment calls?

3. Consequence severity: What happens if the output is bad? Is it wasted time, or is it reputational/financial/legal damage?

Here's how they combine:

LOW-RISK TASKS (legal risk: minimal, accuracy: moderate, consequences: minor)
- Meeting agendas and scheduling
- Internal announcements of non-controversial changes
- FAQ and documentation drafts
- Recruiting email templates
- Brainstorm lists (for you to filter)
- Summary documents of information you provide

Trust level: HIGH. Use AI output with light editing. You're checking for tone and completeness, not accuracy (you know the facts).

MEDIUM-RISK TASKS (legal risk: some, accuracy: high, consequences: moderate)
- Job descriptions
- Policy updates
- Performance review drafts
- Recruiting announcements
- Internal communications about changes
- Benefits communication

Trust level: MEDIUM. Use AI as a starting point, but verify accuracy, check tone, scan for bias. This is heavy editing territory.

HIGH-RISK TASKS (legal risk: significant, accuracy: critical, consequences: severe)
- Termination letters
- Performance improvement plans
- Investigation documentation
- Accommodation responses
- FMLA or ADA-related communication
- Policy language (not updates, the actual policy)
- Severance agreements
- Reference letters

Trust level: LOW. Use AI as a thinking tool, not a drafting tool. Have human legal review everything. Never publish without verification.

Low-Risk: Meeting Agendas and Basic Coordination

The task: You need an agenda for a meeting about the new performance management system. You ask AI to structure the meeting: introductions, overview, Q&A, walkthrough of the new platform, next steps.

Why it's low-risk:
- If the agenda is imperfect, you'll fix it during the meeting.
- Nobody's compliance depends on the agenda.
- The worst outcome is a slightly disorganized meeting.

How to use AI:
- Prompt: "Create an agenda for a 60-minute meeting introducing a new performance management system. Attendees: HR team (5 people), managers (12 people). Focus on getting their buy-in and addressing concerns. Structure: intro (5 min), demo of the platform (20 min), how it affects their job (15 min), Q&A (15 min), next steps (5 min)."
- Trust level: Use as-is or with minor tweaks.

Try This Now: Ask AI to create a meeting agenda for one of your upcoming meetings. Read it. It'll probably be solid. Maybe reorder one section. Send it.

Medium-Risk: Job Descriptions and Policy Updates

The task: You're updating your remote work policy because employees asked for more clarity about approval processes.

Why it's medium-risk:
- Policy language can affect how employees interpret rules.
- If the policy is unclear, employees are frustrated and might make bad decisions.
- It's not high legal risk (assuming you stay within your company's legal boundaries), but it affects operations.

How to use AI:
- Prompt with your actual policy: "Here's our current remote work policy. It's too vague about approval. Can you rewrite it to be clearer, including: how people request remote work, what the approval timeline is, what happens if they work remote without approval? Keep the substance the same, just make it clearer. Tone: warm and clear, not punitive."
- Before sending: Check that you agree with the policy substance (you should. You asked AI to keep it the same). Check that it's clear. Check tone.
- Trust level: Heavy editing. You're verifying accuracy (which should be easy because you provided the source), not just tone.

Another example: Job description

The task: You're hiring a new finance role. You ask AI to draft a job description.

Why it's medium-risk:
- Bad job descriptions get bad candidates or get no candidates.
- Biased language in a job description might legally expose you.
- If you describe requirements incorrectly, you might reject good candidates or miss red flags about bad ones.

How to use AI:
- Prompt: "Draft a job description for a Senior Accountant at a 250-person SaaS company. Reporting to the Controller. The role involves: monthly close, account reconciliation, audit prep, and financial analysis. Our accounting team is two people. The Senior Accountant role is stretched. They'll be hands-on with transactions but also strategic on financial processes. We need someone who can manage ambiguity. We'd want someone with 5-7 years of accounting experience, CPA preferred. Tone: real and specific, not corporate template. Avoid: generic requirements, age language, 'fast-paced environment.' Length: 450-550 words."
- Before sending: Check for bias language (age, gender, ability). Check that requirements are realistic (did you really need 10+ years or is 5-7 enough?). Verify the tone feels like your company.
- Trust level: Medium-high. You're checking for bias and tone. The actual requirements should be accurate because you specified them.

Try This Now: Draft a recruiting email for one open role. Use AI. Scan it against the bias checklist. Does it sound like your company? Then use it.

High-Risk: Terminations, PIPs, and Accommodations

The task: You need to document that an employee is underperforming and needs a PIP.

Why it's high-risk:
- PIP documentation might end up in court if the employee sues for wrongful termination.
- Legal wording matters. A PIP that says "improve performance in the next 30 days or your job is at risk" is different from "we're providing you support and coaching for 30 days." One suggests predetermined outcome; one suggests genuine opportunity.
- If the PIP is vague, you can't demonstrate you gave clear feedback and support.
- If the PIP is discriminatory or retaliatory, it's a lawsuit waiting to happen.

How NOT to use AI:
- Never ask AI: "Draft a PIP for someone who isn't meeting expectations."
- Don't use AI output as-is.
- Don't think "legal review will fix it." Legal review catches major issues but misses problems in tone and framing.

How to use AI correctly:
- Ask AI to think through the structure: "What should a PIP include? What's the difference between a PIP for a performance issue vs. a conduct issue?"
- Use your legal-reviewed PIP template, not AI's draft.
- Use AI to help you write supporting documentation: "Here's what I observed. Help me write this up in clear, factual language without interpretation or emotion."
- Example: You have notes from the past month. AI helps you shape them: "Here's a month of notes about missed deadlines and incomplete work. Structure these factually, just the facts, not judgment, and create a summary of observed performance gaps."
- Before sending: Have legal review it. Every. Single. Time.
- Trust level: Low to none for the final PIP document. Medium for research and thinking tools.

Another high-risk example: Accommodation response

The task: An employee requests to work from home due to a disability. You need to respond.

Why it's high-risk:
- This involves ADA law. Mishandling it is illegal.
- How you respond sets the tone for whether you're seen as accommodating or as fighting the request.
- Every word matters. "We need to evaluate whether this is truly necessary" sounds defensive. "Let's talk about what would help you do your best work" sounds collaborative.

How NOT to use AI:
- Never ask AI: "Draft a response to an accommodation request."
- Don't use AI's legal language without legal review.

How to use AI correctly:
- After your legal team has advised you: Ask AI to help with tone. "Here's what we've decided (we'll approve a 3-day work-from-home arrangement). Help me communicate this in a warm, collaborative tone that shows we're invested in their success."
- Use AI to help you think through logistics: "What should I include in a follow-up email to set up a trial period for remote work?"
- Before sending: Have someone review it who understands ADA. Always.
- Trust level: Low for substantive decisions. Medium-high for tone and logistics once legal has cleared the substance.

Build Your Personal Trust Matrix

You need a simple chart. Here's how:

Column 1: Type of task
- Job descriptions
- Recruiting emails
- Policy updates
- Performance reviews
- Onboarding checklists
- Meeting agendas
- etc.

Column 2: Legal risk
- None / Low / Medium / High

Column 3: Accuracy risk
- Low / Medium / High

Column 4: Trust level
- High (use with light editing) / Medium (heavy editing) / Low (thinking tool only, never use output as-is)

Column 5: Next step
- Use as-is or minimal edits
- Heavy human editing before use
- Legal review required before use
- Don't use AI output, use AI for brainstorm only

Example:

Task
Legal Risk
Accuracy Risk
Trust Level
Action

Job descriptions
Medium
High
Medium
AI draft + bias check + tone review

Meeting agendas
None
Low
High
Use with minimal edits

Performance reviews
Low
Medium
Medium
AI draft + specific examples + no personality language

Termination letters
High
High
Low
AI for thinking, human drafts + legal review

Recruiting emails
Low
Low
High
Use as-is or light edits

Policy updates
Medium
High
Medium
AI rewrite of existing policy + verify substance

Accommodation responses
High
High
Low
Legal decides substance, AI helps tone only

Investigation summaries
High
High
Low
Thinking tool only + legal review mandatory

Build yours. Keep it. Use it every time you're about to hand something to AI.

Tip: When in doubt about risk level, ask: "Could this end up in court?" If yes, it's high-risk. Assume legal review is required.

The Override Decision

Sometimes AI output is good enough for low-risk tasks, but you still need to override it.

When to override:
- Tone is wrong for your company
- Information is inaccurate
- Something's missing that employees need to know
- The output doesn't match your voice or values
- You wouldn't personally send this

When NOT to override:
- It's "not quite what I imagined" (editing is different from overriding)
- You'd prefer different wording but both are fine (trust the AI output)
- It's close enough (remember, you're saving time)

Example: Recruiting email

AI draft: "We're excited about your background in product management and think you'd be a great fit for our Director of Product role. Let's set up a time to chat."

You're tempted to override: "This doesn't mention why we're reaching out or what makes us different from other companies."

Should you override? Maybe. If the version without those details gets 0% response rate, you're losing candidates. If it gets 50% response rate, you're fine. It's just a first touchpoint.

The deciding question: "Is this good enough to get the result I need?" If yes, send it. You're not writing for perfection; you're writing to move things forward.

Try This Now: Build Your Matrix

• 
  List your top 10 recurring HR tasks (job descriptions, recruiting emails, policy updates, performance reviews, onboarding, etc.).
• 
  For each task, rate the risk:
  - Legal risk: None / Low / Medium / High (would a lawyer care about this?)
  - Accuracy risk: Low / Medium / High (how much does correctness matter?)
• 
  Determine your trust level based on combined risk.
• 
  Decide your action for each task:
  - "I'll use AI output with light editing"
  - "I'll use AI as a starting point and heavily edit"
  - "I'll use AI for brainstorming/thinking, never use output as-is"
  - "I won't use AI for this task"
• 
  Save this as a reference. When you're about to hand a task to AI, check your matrix. It tells you how careful you need to be.

Practical Application - "What to Do Monday Morning"

• 
  Identify one high-risk task you currently do (probably terminations, accommodations, or investigation-related).
• 
  Ask yourself: "Am I using AI for this?" If yes, stop. If no, you're good.
• 
  Identify one medium-risk task you do regularly (probably job descriptions, policy updates, or recruiting).
• 
  Audit your current approach. Are you sending output with light editing, or heavy editing? Should you shift?
• 
  Create your trust matrix this week. Keep it visible. Use it every time.
• 
  For high-risk tasks, make sure legal review is part of your process. Not optional. Built in.

Key Takeaways

• Risk determines trust: Low-risk tasks = light editing. Medium-risk = heavy editing. High-risk = legal review, never use output as-is.
  - Three variables matter: Legal risk, accuracy requirements, consequence severity.
  - Build your personal matrix: Know which tasks are safe to delegate to AI and which need human oversight.
  - For high-risk tasks, legal review is non-negotiable: PIPs, terminations, accommodations, investigations. Every time.
  - Override when tone, accuracy, or values are wrong: But don't over-edit. Sometimes "good enough" is actually good enough.

FAQ

Q: How do I know if something is high-risk?
A: Ask: "Could this end up in court? Could someone be harmed by wrong information?" If yes to either, it's high-risk.

Q: What if I don't have a legal team to review?
A: For high-risk tasks, you need external legal review. This isn't optional. Budget for it. It's cheaper than lawsuits.

Q: Can I use AI for anything related to accommodations?
A: You can use AI to think through logistics, but the decision and language must be reviewed by someone who understands ADA law. The stakes are too high.

Q: What about performance reviews? Are they high-risk?
A: Medium-risk. Reviews affect promotions and terminations, but the risk is about quality (being specific, not discriminatory) not legal landmines like PIPs. Check for bias and specificity; you don't necessarily need legal review.

Q: Should I always have legal review everything?
A: No. That's inefficient and expensive. But yes for anything touching employment law, accommodations, investigations, or terminations. Your matrix will tell you.

What's Next

You've now got the foundations: how to prompt, how to recognize useful output, how to spot red flags, and how to know when to trust it. Chapter 2 shifts into specific HR function: recruiting. You'll learn how to use AI to write better job descriptions, source candidates, screen resumes, and prepare interviews.